signspell 2.0.0

Live ASL fingerspelling alphabet recognition with a polished webcam UI.

signspell

Live ASL fingerspelling alphabet recognition — straight from your webcam.

signspell recognises the American Sign Language manual alphabet (A–Z) in real time using MediaPipe hand tracking and an LSTM model trained on 30-frame keypoint sequences. It ships with a pretrained model and a polished webcam UI, and it works both as a command-line tool and an importable library.

The American Sign Language fingerspelling alphabet that signspell recognises.

---

Install

pip install signspell

Requires Python 3.9–3.12. A webcam is required for live recognition.

Run it (CLI)

signspell                  # default webcam, bundled model, pro UI
signspell --camera 1       # pick a different camera
signspell --threshold 0.6  # require higher confidence
signspell --no-mirror      # disable mirrored view

In-window keys: q quit · c clear sentence · SPACE add a space.

Use it (library)

Run the full UI from code:

import signspell
signspell.run()

Or drive recognition yourself, frame by frame:

import cv2
import signspell

rec = signspell.Recognizer()
cap = cv2.VideoCapture(0)

while True:
    ok, frame = cap.read()
    if not ok:
        break
    letter, confidence, probs = rec.predict(frame)
    if letter:
        print(letter, f"{confidence:.2f}")

The Recognizer keeps a rolling 30-frame buffer internally, so you just feed frames and read predictions. letter is None until the buffer fills or when confidence is below the threshold.

How it works

1. MediaPipe Holistic extracts 21 right-hand landmarks per frame.
2. The last 30 frames of (x, y, z) keypoints form a sequence.
3. An LSTM classifies the sequence into one of 26 letters.
4. A short stability window prevents flicker before a letter is committed.

Bring your own model

signspell --model path/to/your_model.h5

signspell.Recognizer(model_path="path/to/your_model.h5")

Your model must accept input shape (1, 30, 63) and output 26 class scores.

License

MIT © Sundar